Due to an error during an update to the system kernel on a database
server backing a web application, we ran for a month (mid-December
to mid-January) with the WAL files in the pg_xlog subdirectory on
the same 40-spindle array as the data -- only the OS was on a
separate mirrored pair of drives. When we found the problem, we
moved the pg_xlog directory back to its own mirrored pair of drives
and symlinked to it. It's a pretty dramatic difference.
Graph attached, this is response time for a URL request (like what a
user from the Internet would issue) which runs 15 queries and
formats the results. Response time is 85 ms without putting WAL on
its own RAID; 50 ms with it on its own RAID. This is a real-world,
production web site with 1.3 TB data, millions of hits per day, and
it's an active replication target 24/7.
Frankly, I was quite surprised by this, since some of the benchmarks
people have published on the effects of using a separate RAID for
the WAL files have only shown a one or two percent difference when
using a hardware RAID controller with BBU cache configured for
write-back.
No assistance needed; just posting a performance data point.
-Kevin
